Working note — not peer reviewed Working paper v0.2

Interruption Budgets: Rationing Human Attention Across Concurrent AI Agents

Authors
DataKnife Research
Date
Area
Attention-aware interfaces

Abstract

Interfaces for supervising multiple AI agents inherit a notification model built for single applications: every event is either silent or an interruption. We describe why this model collapses under agent concurrency, and propose an alternative framing — an interruption budget — in which agents spend from a bounded attention allowance and interfaces render most state peripherally instead of interruptively. We outline the design space, the signal taxonomy we are using in the Ambient Agent Dashboard prototype, and the evaluation we intend to run against it.

What we investigated

When one person supervises many long-running AI agents, what should the interface interrupt them for? Current tooling answers implicitly: either everything (notification per event) or nothing (check the transcript yourself). We investigated the structure of that failure and what a deliberate attention model would look like instead.

Why it matters

Agent concurrency is becoming normal. A person running five agents receives the interruption load of five chat applications, but the useful information rate is far lower — most agent events require no human action. Interfaces that cannot distinguish “needs a human” from “produced an event” will cap how many agents one person can effectively run, regardless of how capable the agents are.

Method

This is design research grounded in prototype instrumentation, not a controlled study. We are (1) cataloging agent events from our own daily multi-agent use into a signal taxonomy — progress, state change, blocked-on-resource, blocked-on-human, completion, failure; (2) implementing competing presentation policies for each class in the Ambient Agent Dashboard prototype; and (3) logging interruption events and human response behavior in our own use to compare policies. A structured evaluation with outside users is planned but has not run.

Findings

Stated as working design findings from prototype use, not validated results:

  • Most agent events are peripheral by nature. In our own usage logs, the overwhelming majority of events require no action; rendering them as anything louder than ambient state is pure attention cost.
  • “Blocked on human” is the only class that clearly earns interruption — and it benefits from stating what it needs in the interruption itself, not in the transcript behind the click.
  • A budget framing changes design decisions. Asking “is this worth spending interruption budget on?” produced different, and in our use better, defaults than asking “should this notify?”

Limitations

This note reports the experience of the people building the system — the most biased possible population. Event taxonomies drawn from our own agents may not transfer to other workloads. No controlled evaluation has been run; response-behavior logging is in-house only. These findings should be read as hypotheses the prototype exists to test, and this note will be revised as the evaluation produces evidence — including if it produces negative results.

Product implications

The Ambient Agent Dashboard’s signal hierarchy — ambient tile state, then audio cue, then interruption — is a direct implementation of the budget model. If the evaluation holds up, the same model applies anywhere one human supervises many autonomous processes; if it does not, the peripheral-first rendering is still independently defensible on interruption-cost grounds.